Claude adopts Anthropic's thinking reduction, boosting AI efficiency today

Claude’s engineering output has taken a hard left turn since Anthropic quietly flipped the switch on its “thinking reduction” feature in early March. According to a Gist analysis of 6,852 Claude code‑session files, the rollout erased 17,871 thinking blocks and 234,760 tool calls, effectively stripping the model of the internal monologue it used to plan multi‑step research and careful code edits. The data show a stark timeline: thinking was fully visible through March 4, then fell to 1.5 % on March 5, 24.7 % on March 7, 58.4 % on March 8, and hit 100 % redaction by March 12. That week‑long staged deployment coincides exactly with the quality regression users reported on March 8, when the redaction crossed the 50 % threshold.

Even before the redaction, Claude’s “thinking depth” was already on a downhill slope. Gist’s correlation analysis—using a 0.971 Pearson link between a signature field and content length—estimates median thinking token length at roughly 2,200 characters in late January, dropping to about 720 characters by late February (a 67 % decline). By early March the median had slipped to roughly 560 characters, a 75 % reduction from the baseline. The redaction rollout simply made that loss invisible to end users, cloaking a trend that was already eroding the model’s capacity for sustained reasoning.

The behavioral fallout is measurable. Gist computed quality metrics on more than 18,000 user prompts before the thinking cutback and again after March 8. “Stop hook” violations—automated guards that catch ownership‑dodging, premature stopping, and permission‑seeking—spiked from zero to 173 instances in just 17 days, averaging ten per day. Frustration signals in user prompts rose from 5.8 % to 9.8 % (+68 %). Corrections needed for “ownership‑dodging” doubled, climbing from six to thirteen cases (+117 %). Meanwhile, prompts per session fell 22 % (35.9 → 27.9), and sessions that featured five or more reasoning loops jumped from none to seven, indicating that Claude was now more likely to abort complex chains of thought.

Perhaps the most telling symptom is the shift in tool usage. Gist’s audit of 234,760 tool invocations reveals a dramatic swing from a research‑first to an edit‑first workflow. During the “good” period (Jan 30 – Feb 12), Claude performed 6.6 file reads for every edit, with reads accounting for 46.5 % of tool calls and edits a modest 7.1 %. By the “degraded” period (Mar 8 – Mar 23), reads per edit collapsed to 2.0, a 70 % reduction, while edit calls rose to 15.4 % of total tool usage. In other words, the model stopped scanning the codebase before making changes, leading to the sloppy, low‑quality patches users have been flagging.

Anthropic’s own internal memo, referenced by the Gist report, frames the thinking blocks as “load‑bearing” for senior‑engineering workflows. The analysis suggests that extended thinking tokens are not a luxury but a structural necessity for multi‑step research, convention adherence, and careful code modification. When those tokens are stripped away, Claude’s behavior pivots toward quick, surface‑level edits, sacrificing the depth that power users rely on. The data points to a clear trade‑off: cutting thinking tokens boosts raw throughput but erodes the nuanced reasoning that makes Claude valuable for complex engineering tasks. As the community digests these findings, the pressure mounts on Anthropic to recalibrate token allocation—or risk alienating the very senior engineers who once championed Claude as a productivity booster.